The AI industry has invented a new language faster than most people can learn it, and even seasoned tech professionals are feeling left behind. Terms like LLM (large language model), RAG (retrieval-augmented generation), RLHF (reinforcement learning from human feedback), and dozens of others now dominate product meetings, investor pitches, and industry panels. To address this growing vocabulary crisis, TechCrunch published an updated glossary in early July 2026 that attempts to translate AI's increasingly complex terminology into plain English.

The timing couldn't be more critical. As AI systems like DeepSeek-R1 and other reasoning models gain prominence, the concepts underlying them have become essential knowledge for anyone building with, investing in, or simply trying to keep pace with artificial intelligence. Yet even the world's leading AI researchers can't agree on what some of the most fundamental terms actually mean.

What Exactly Is AGI, and Why Can't Experts Agree?

Artificial General Intelligence, or AGI, sits at the center of this definitional chaos. The term generally refers to AI systems that are more capable than the average human at many, if not most, tasks. But beyond that broad description, consensus collapses.

OpenAI CEO Sam Altman has described AGI as the "equivalent of a median human that you could hire as a co-worker." OpenAI's official charter takes a slightly different approach, defining AGI as "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind, meanwhile, views AGI as "AI that is at least as capable as humans at most cognitive tasks". These definitions sound similar on the surface, but the nuances matter enormously for research priorities, safety considerations, and investment decisions.

"Even experts disagree, making this one of the most debated concepts in the field," noted the TechCrunch glossary authors.

TechCrunch Editorial Team

This definitional ambiguity has real consequences. When investors hear a startup claim it's "approaching AGI," they're interpreting that claim through different mental models depending on which definition they've internalized. The same applies to researchers publishing papers or policymakers drafting regulations.

How Are AI Systems Actually Learning to Reason Step by Step?

One of the most significant developments in modern AI is chain-of-thought reasoning, a technique that has become central to how advanced models like DeepSeek-R1 operate. Unlike traditional language models that produce answers directly, chain-of-thought reasoning breaks down complex problems into intermediate steps, much like a human working through a math problem with pen and paper.

Consider a classic logic puzzle: "A farmer has chickens and cows. Together they have 40 heads and 120 legs. How many of each animal does the farmer have?" A human would likely write out equations to solve this. An AI model using chain-of-thought reasoning does something similar, showing its work step by step. This approach produces significantly more accurate answers, especially in logic and coding contexts, though it typically takes longer to generate a response.

Reasoning models are developed from traditional large language models and optimized specifically for this structured thinking through reinforcement learning, a training technique that rewards correct intermediate steps and final answers. This is why models built around chain-of-thought reasoning have become so valuable for tasks requiring genuine problem-solving rather than pattern matching.

Steps to Understanding AI's Emerging Terminology

• Start with foundational concepts: Begin by understanding what large language models (LLMs) are and how they work, as nearly every other AI term builds on this foundation. LLMs are AI systems trained on vast amounts of text data to predict and generate human language.
• Learn the reasoning techniques: Familiarize yourself with chain-of-thought reasoning and how it differs from traditional language model outputs. This distinction is crucial for evaluating claims about AI capabilities in problem-solving domains.
• Understand specialized applications: Recognize that terms like "coding agent" and "AI agent" refer to specialized versions of AI systems designed for specific tasks, not general-purpose systems that can do everything.
• Stay updated on evolving definitions: Accept that AI terminology is still being defined and refined. What "AGI" means today may shift as the field progresses, so treat glossaries as living documents rather than final authorities.

What Are AI Agents, and How Do They Differ From Chatbots?

An AI agent represents a significant step beyond traditional chatbots. While a chatbot responds to user queries, an AI agent uses AI technologies to perform a series of tasks autonomously on your behalf, going far beyond simple question-and-answer interactions. These agents can file expenses, book restaurant reservations, arrange travel, or even write and maintain code without constant human direction.

The distinction matters because it signals a shift in how AI systems interact with the world. Rather than simply providing information, agents can take action. However, the infrastructure needed to fully deliver on these capabilities is still being built out, and the term "AI agent" can mean different things to different people depending on the context and sophistication level.

A specialized subset called coding agents applies this autonomous capability specifically to software development. Rather than suggesting code snippets for a human to review and manually integrate, a coding agent can write, test, debug, and deploy code autonomously across entire codebases. It spots bugs, runs tests, and pushes fixes with minimal human oversight. As one glossary description puts it, it's "like hiring a very fast intern who never sleeps and never loses focus," though human review of the work remains essential.

Why Are Distillation and Knowledge Transfer Becoming Controversial?

Distillation is a technique that extracts knowledge from a large AI model, called the "teacher," to train a smaller, more efficient model, called the "student." Developers send requests to the teacher model, record its outputs, and use those outputs to train the student model to approximate the teacher's behavior. This approach allows companies to create faster, cheaper versions of their models with minimal performance loss.

OpenAI likely used distillation internally to develop GPT-4 Turbo, a faster version of its original GPT-4 model. The technique is standard practice across the AI industry. However, distillation has become controversial because some AI companies have allegedly used it to copy competitor models, which violates the terms of service of most AI APIs and chat assistants. This has raised questions about intellectual property protection in the AI space and whether current legal frameworks adequately address knowledge transfer between competing systems.

The broader context here is that as AI models become more valuable and competitive pressures intensify, companies are exploring every technical avenue to catch up with frontier models. Distillation sits in a gray zone where the technique itself is legitimate, but its application can cross ethical and legal lines depending on whose model is being distilled and how.

What Other Core Concepts Should You Know?

Beyond AGI, chain-of-thought reasoning, and agents, the updated glossary covers several other essential terms that appear frequently in AI discussions:

• Diffusion: The technology powering many image, music, and text generation models. Inspired by physics, diffusion systems gradually add noise to data until its structure is destroyed, then learn to reverse this process to reconstruct original data from noise.
• Deep learning: A subset of machine learning where AI algorithms use multi-layered artificial neural networks inspired by the human brain. These systems can identify important patterns in data without humans explicitly defining those features.
• Fine-tuning: The further training of an AI model to optimize performance for specific tasks or domains, typically by feeding in new, specialized data. Many AI startups use this approach to adapt large language models for their particular industry or use case.
• Compute: The computational power that fuels AI systems, including the hardware like GPUs (graphics processing units), CPUs (central processing units), and TPUs (tensor processing units) that form the infrastructure backbone of the modern AI industry.

The fact that TechCrunch felt compelled to publish and regularly update a comprehensive AI glossary underscores a fundamental challenge facing the industry. As AI capabilities expand and new techniques emerge, the vocabulary required to discuss them intelligently grows faster than most people can absorb. For professionals building with these technologies, investors evaluating AI companies, and policymakers crafting regulations, having a shared understanding of core terminology isn't just helpful; it's essential for making informed decisions in a field that's moving at unprecedented speed.